toward an aposynthesis of topic continuity and ... · toward an aposynthesis of topic continuity...

c© 2002 Association for Computational Linguistics

Toward an Aposynthesis of TopicContinuity and Intrasentential Anaphora

Eleni Miltsakaki∗University of Pennsylvania

The problem of proposing referents for anaphoric expressions has been extensively researched inthe literature and significant insights have been gained through the various approaches. However,no single model is capable of handling all the cases. We argue that this is due to a failure of themodels to identify two distinct processes. Drawing on current insights and empirical data fromvarious languages we propose an aposynthetic1 model of discourse in which topic continuity,computed across units, and focusing preferences internal to these units are subject to differentmechanisms. The observed focusing preferences across the units (i.e., intersententially) are bestmodeled structurally, along the lines suggested in centering theory. The focusing mechanismwithin the unit is subject to preferences projected by the semantics of the verbs and the connectivesin the unit as suggested in semantic/pragmatic focusing accounts. We show that this distinctionnot only overcomes important problems in anaphora resolution but also reconciles seeminglycontradictory experimental results reported in the literature. We specify a model of anaphoraresolution that interleaves the two mechanisms. We test the central hypotheses of the proposedmodel with an experimental study in English and a corpus-based study in Greek.

1. The Problem

Extensive research reported in the anaphora resolution literature has focused on theproblem of proposing referents for pronominals.2 First, centering, formulated as amodel of the relationship between attentional state and form of referring expressions,was utilized as the basis of an algorithm for binding pronominals on the intersen-tential level (Brennan, Walker-Friedman, and Pollard 1987). The proposed algorithm(henceforth the BFP algorithm) gives the correct interpretation for the pronominal hein example (1), stating a preference to resolve the pronominal to Max rather than Fred.

(1) a. Max is waiting for Fred.

b. He invited him for dinner.

It was soon observed, however, that the BFP algorithm was not capable of handlingcases of intrasentential anaphora such as in (2) (adapted from Suri, McCoy, and De-Cristofaro [1999]).

∗ Institute of Research in Cognitive Science, Philadelphia, PA 19104. E-mail: [email protected] “Aposynthesis” is a Greek word that means “decomposition,” that is, pulling apart the components

that constitute what appears to be a uniform entity.2 Although a significant amount of research in anaphora resolution has been carried out in statistical

approaches, reviewing such approaches is well beyond the scope of the current article.

320

Computational Linguistics Volume 28, Number 3

(2) a. Dodge was robbed by an ex-convict.

b. The ex-convict tied him up

c. because he wasn’t cooperating.

d. Then he took all the money and ran.

The centering-based BFP algorithm would have a preference to resolve he in (2d) toDodge and not to the ex-convict, based on a preference for a Continue transition.

Alternative approaches to anaphora resolution have sought to account for the re-solution facts by proposing a semantic/pragmatic rather than structural mechanism.Stevenson et al. (2000) argue that both verbs and connectives have focusing proper-ties affecting the preferred interpretation of pronominals. So in (3), the verb focusinghighlights Bill, since Bill is the person associated with the endpoint of the event of crit-icizing. The connective, so, directs attention to the consequences and hence reinforcesthe focus on Bill.

(3) a. John criticized Bill,

b. so he tried to correct the fault.

The semantic/pragmatic focusing account runs into the type of problem demon-strated in (4), where the preferred interpretation for he is John, that is, the structuralsubject, independent of semantic/pragmatic factors.3 In such discourses it seems thata structural account is at play (in the sense of Grosz and Sidner [1986]).

(4) a. John criticized Bill.

b. Next, he insulted Susan.

This article sets out to explicate the behavior of pronominals demonstrated in theabove examples. Gaining significant insights from current research in anaphora reso-lution, we reconcile what seem to be contradictory findings in a model according towhich inter- and intrasentential anaphora are not subject to the same mechanism. Weargue that the shortcomings of the proposed algorithms are due to confounding twodistinct processes, namely, topic continuity and the internal structure of the sentence.4

We conclude that intersentential anaphora is subject to structural constraints, whereasintrasentential anaphora is subject to grammatical as well as semantic/pragmatic con-straints. We define the notion of discourse unit and propose a two-level approach toanaphora resolution. Within the unit, anaphora resolution is performed locally and isconstrained by the grammar and semantic properties of the predicates and the subor-dinate conjunctions. This process outputs unresolved anaphoric expressions for whichpotential referents are picked from a centering-style ranked list of entities constructedin the previous unit.

3 Experimental results regarding these cases are reported in Stevenson et al. (2000).4 We use the term “topic” to describe a centered entity, that is, the entity that the discourse is “about.”

The notion of a centered entity is a discourse construct distinct from “topic” or “theme” as defined ininformation structure. Elsewhere we have introduced the term “attention structure in discourse” todescribe mechanisms, linguistic or nonlinguistic, that language users employ to navigate the hearer’sattention in discourse. Topic continuity is derivative of attention structure in discourse. In this article,however, we have opted for the more transparent term “topic continuity,” as it describes thephenomenon we are mostly concerned with in pretheoretical terms.

321

Miltsakaki Topic Continuity and Intrasentential Anaphora

The article is organized as follows. In Section 2 we give a brief overview ofcentering-based models of anaphora resolution, discuss their shortcomings, and con-trast them with the semantic-focusing account suggested in Stevenson et al. (2000). InSection 3, we present the discourse model we adopt and the specifications we proposefor anaphora resolution across and within centering update units. In Section 4, we testthe central hypotheses of the proposed model in two studies: an experimental studyin English and a corpus-based study in Greek. We conclude with a general discussionin Section 5.

2. Issues and Insights in Anaphora Resolution

2.1 The BFP AlgorithmBrennan, Walker-Friedman, and Pollard (1987) were the first to use the centering modelas the basis for an anaphora resolution algorithm. The centering model (Grosz andSidner 1986; Grosz, Joshi, and Weinstein 1983) makes the following assumptions:

1. A discourse segment consists of a sequence of utterances, U1 , . . . , Un.

2. For each utterance, a ranked list of evoked discourse entities isconstructed, designated as the Cf list.

3. The highest-ranked element of the Cf list is called the preferred center(Cp).

4. The highest-ranked entity in the Cf list of Ui−1 realized in Ui is thebackward-looking center (Cb).

There are several types of topic transitions from one utterance to the next dependingon whether the Cb is retained over two consecutive utterances Un−1 and Un andwhether this Cb is also the Cp of Un (see Table 1). The distinction between a SmoothShift and a Rough Shift is due to Brennan, Walker-Friedman, and Pollard (1987), whoobserved that the centering model generates ambiguity in cases such as (5):

(5) a. Brennan drives an Alfa Romeo.

b. She drives too fast.

c. Friedman races her on weekends.

d. She often beats her.

Adding weight to the status of the Cp in (5c) makes it possible to resolve the pronom-inal she in (5d) successfully to Friedman. We will return to the issue of ambiguityshortly.

The BFP algorithm consists of three basic steps:

1. Generate possible Cb-Cf combinations.

2. Filter by constraints (e.g., contra-indexing, sortal predicates, centeringrules and constraints).

3. Rank by transition orderings (Continue > Retain > Smooth Shift >Rough Shift).

322


Table 1Centering transitions.

Cb(Ui) = Cb(Ui−1 ) Cb(Ui) �= Cb(Ui−1 )Cb(Ui) = Cp(Ui) Continue Smooth ShiftCb(Ui) �= Cp(Ui) Retain Rough Shift

Some of the shortcomings of the BFP algorithm are discussed by Prasad and Strube(2000), who observe that it makes two strategic errors. Prasad and Strube’s observa-tions are made with respect to Hindi but hold in English and Modern Greek, as shownin (6) and (7), respectively.

The first of these errors occurs in cases in which Cb(Ui−1 ) is different fromCp(Ui−1 ). In such cases, the preference for a Continue transition is responsible forthe pronominal in Ui being resolved to the Cb(Ui−1 ) and not to the Cp(Ui−1 ).

(6) a. Elleni saw Maryj at school.

b. Maryj didn’t talk to heri.

c. Shej took herj friends and walked away.

(7) a. I Elenii ide ti Mariaj sto sholio.the Eleni saw the Maria at-the school.Elenii saw Mariaj at school.

b. I Mariaj den tisi milise.the Maria not to-her talked.Mariaj didn’t talk to heri.

c. NULLj pire tis files tisj ki NULLj efige.NULL took the friends her and NULL left.Shej took herj friends and left.

There is an important observation to be made here, which we present as the firstindication for the distinction between topic continuity and anaphora resolution. Onthe one hand, the BFP centering-based algorithm makes a resolution error opting fora Continue transition in (6c) and (7c). On the other hand, anaphora aside, the topictransition identified by centering is intuitively correct. In (6) and (7), the discourse isinitiated with Ellen as the current topic, Maria is introduced as an entity related to thecurrent topic, and then the discourse shifts to Maria to elaborate on her doings. Theshift is in fact anticipated by the promotion of Maria from the object position in (6a)and (7a) to the subject position in (6b) and (7b).

The second error observed by Prasad and Strube (2000) is that the BFP algorithmgenerates ambiguity when Ui−1 is discourse initial. Example (8) is given as illustration.

(8) a. John gave a lot of his property to George.

b. His current salary exceeded the average salary by a lot.

Given that the Cb in the discourse initial (8a) is unspecified, Continue transitions aregenerated when resolving his to either John or George. At this point, the BFP algorithmis not capable of reaching a decision.

323


The solution we propose for the two problems is simple: the preferred antecedentfor the pronominal in Ui is the highest-ranked entity in Ui−1 that is compatible withthe anaphoric expression. Compatibility is defined in terms of agreement features(number and gender in the case of English). The proposed solution is consistent withthe centering model. The most relevant centering notion for anaphora resolution isthe pronoun rule, which stipulates that if an entity is realized as a pronoun, then sois the Cb. Opting for resolution to the highest-ranked entity in the previous entity isprecisely supported by the pronoun rule because the highest-ranked entity realizedin the following utterance is the Cb. On the other hand, using centering transitionsfor anaphora resolution does not necessarily follow from the original formulation ofcentering. Centering transitions, as originally formulated and as confirmed by the datadiscussed above, are best at identifying degrees of topic continuity. There is no a pri-ori reason to expect that they will perform equally well in identifying pronominalreferents. This is because assuming maximal coherence (preference for Continue tran-sitions) overlooks properties of attention structure in discourse: strategies that hearersuse to signal attention shifts to new centers while maintaining coherence. A SmoothShift may be intended and signaled appropriately by, for example, promoting a propername from object to subject position. Interpreting pronominals in accordance with thepronoun rule as suggested here exploits precisely such strategies.

We conclude from this section that although centering transitions successfullyidentify topic continuity in discourse, in anaphora resolution the most useful centeringnotion is not the transitions themselves but the Cf list ranking in combination withthe pronoun rule.

2.2 Functional CenteringStrube and Hahn (1996, 1999) elaborate on the nature of the Cf list and propose acentering-based model of anaphora resolution in which the Cf ranking is based noton grammatical function but on functional information status. They recast centeringnotions in terms of Danes’s (1974) trichotomy between given information, theme, andnew information. The Cb(Ui), the most highly ranked element of Cf(Ui−1 ) realized inUi, corresponds to the element that represents given information. The Cp(Ui) corre-sponds to the theme of Ui. The rhematic elements of Ui are the ones not containedin Ui−1 . Although the original motivation for the functional recast of centering wasdue to German, a free-word-order language, Strube and Hahn (1996) claim that thefunctional framework is superior because fixed- and free-word-order languages canbe accounted for using the same principles. They argue against Walker, Iida, and Cote(1994), in which the Cf ranking is viewed as a language-specific parameter that needsto be set.

In what follows we will remain agnostic to the suitability of the functional center-ing framework for German. We will show, however, that functional centering is notthe appropriate framework for all free-word-order languages, much less for languagesuniversally. We bring in evidence from Modern Greek, a free-word-order language.

To identify the factors determining Cf ranking in Greek, we employ Rambow’s(1993) diagnostic5 to test whether surface word order or grammatical function is themost reliable indicator of salience. The relevant examples for the Greek version of

5 Rambow suggests that the order of entities in the position between finite and nonfinite verbs inGerman (Mittelfeld) affects their salience. Gender in German is grammaticized, so Rambow constructsan example with two same-gender entities in Mittelfeld and uses an ambiguous pronoun in subsequentdiscourse to determine which of the two entities is more salient. The constructed example is givenbelow. “Fem” indicates that the noun phrases are gender marked “feminine.”

324


Rambow’s diagnostic are shown in (9) and (10). The null pronominal in (9b) and (10b)resolves to the subject irrespective of its surface position.6 The relevant indicator ofsalience in the Cf list appears to be grammatical function, at least subjecthood.7

(9) a. I prosfati diefthetisii veltioni tin ikonomiki politikij?the recent arrangement improve the economic policy?Does the recent arrangementi improve the economic policyj?

b. Ohi, (nulli) ine aneparkis.No, (it) is inadequate.No, iti is inadequate.

(10) a. Tin ikonomiki politikij tij veltioni i prosfati diefthetisii?the economic policy it-(clitic) improve the recent arrangement?Does the recent arrangementi improve the economic policyj?

b. Ohi, (nulli) ine aneparkis.No, (it) is inadequate.No, iti is inadequate.

Further evidence for the role of grammatical function in Greek comes from syntac-tic objects.8 In Greek (and also in Turkish), a strong pronominal or a full noun phrase(NP) must be used to promote the object of Ui−1 to the subject position of Ui.9 As theinfelicitous interpretations (indicated by the pound sign) show in (11b), reference tothe object Yorgo becomes felicitous only with the use of name repetition or a strong

(1) a. Glauben Sie, dass [eine solche Massnahme]i [der russischen Wirtshaft]j helfen kann?think you that a such measure-Fem the Russian economy-Fem help can?Do you think that such a measure can help the Russian economy?

b. Nein, siei ist viel zu primitiv.no, she is much too primitive.No, it’s much too primitive.

6 Gender and lexical considerations are controlled. Both economical policy and arrangement are feminineand they can both be inadequate. Also, we have presented the diagnostic test and confirmed thejudgment with a sizable group of native speakers of Greek attending the 15th International Symposiumon Theoretical and Applied Linguistics (Miltsakaki 2001).

7 It is interesting that in Turkish, another free-word-order language, it has also been shown (Turan 1998)that the strongest indicator of salience is subjecthood.

8 Greek has two pronominal systems: weak pronouns that must cliticize to the verb and strong pronounsthat are syntactically similar to full NPs. Dropped subjects are considered part of the system of weakpronouns. In Miltsakaki (2000), we argue that speakers of various languages use available nominal andpronominal forms and prosodic features in spoken language to signal attention structure in discourse.Greek speakers with a three-way distinction in their nominal system (i.e., full noun phrases and weakand strong pronominals) use strong pronominals to signal reference to an entity previously evoked indiscourse that is not, however, the most salient entity. This use of strong pronominals is equivalent tocertain prosodic effects in English. As noted by an anonymous reviewer, there is extensive literature onthe effects of prosody in pronominal interpretation. For example, it has been observed that prominentstress on the pronominals in (1) yields cospecification of he with Bill and him with John.

(1) John criticized Bill. Then, HE criticized HIM.

The need to recruit special prosody to achieve resolution to Bill indicates that structural focusing isindeed at work projecting strong “default” focusing preferences. In (1), there is sufficient semanticinformation to help the hearer arrive at the intended interpretation. If there was no defaultinterpretation available at hand there would be no need to evoke prosodic effects. Once the linguisticencoding of speakers’ strategies for building attention structure in discourse are identified,incorporating them in the centering framework should be trivial.

9 A “full NP” is any noun phrase that contains a head noun, either common or proper.

325


pronominal, shown in (11c) and (11d).10 We take this as further evidence that objectsrank lower than subjects in Greek.

(11) O Yannisi proskalese ton Yorgoj.the John invited the Yorgo.Johni invited Georgej.

a. nulli tuj prosfere ena poto.he him offered a drink.Hei offered himj a drink.

b. #null j #tu i prosfere ena poto.he him offered a drink.Hej offered himi a drink.

c. O Yorgosi tui prosfere ena poto.the George him offered a drink.George offered himi a drink.

d. Ekinosj tui prosfere ena poto.he-strong him offered a drink.HEj offered himi a drink.

Finally, to test the current results against the functional centering alternative, wereplace the definite subject in (9) with an indefinite noun phrase. As shown in (12),the subject is an indefinite noun phrase representing new (or hearer-new) informationand the object is a definite phrase, encoding old (or hearer-old) information. The nullpronominal in (12b) resolves to the subject of (12a), and the information status of thepotential antecedents is disregarded.

(12) a. Mia kainourgia diefthetisii tha veltiosi tin ikonomiki politikij?a new arrangement will improve the economic policy?Will a new arrangementi improve the economic policyj?

b. Ohi, (nulli) tha ine aneparkis.No, (it) will be inadequate.No, it will be inadequate.

That the information status is not the relevant factor in discourse salience, at leastnot cross-linguistically, is also confirmed in Turan (1998) for Turkish and in Prasad andStrube (2000) for Hindi. In both of these languages, the relevant factor for the rankingof elements in the Cf list is grammatical function.

We conclude that information status (or hearer status) is not universally the mostimportant factor determining discourse salience (in Cf ranking). Given the facts ofpronominalization, we maintain that, at least for English, Greek, Hindi, and Turkish,grammatical function can most reliably determine the relative salience of entities.

10 Empirical evidence for the use of strong pronominals to signal reference to nonsalient entities in Greekis provided in Dimitriadis (1996). Further functions of strong pronominals in Greek are identified inMiltsakaki (1999, 2001).

326


2.3 The S-list AlgorithmA further modification of the centering model is proposed by Strube (1998), who re-places the functions of the backward-looking center and the centering transitions withthe ordering among elements of what he calls the S-list, that is, the list of salientdiscourse entities. The S-list ranking criteria define a preference for hearer-old overhearer-new discourse entities and are intended to reflect the attentional state of thehearer at any given point in discourse processing. The S-list is generated incremen-tally and is updated every time an anaphoric element is resolved. Anaphoric elementsare resolved with a lookup in the S-list. The elements of the S-list are looked up forcompatibility in the order determined by their information status ranking (hearer-oldentities are looked up before hearer-new entities). When the analysis of the utteranceis finished (processed left to right), the discourse entities that are not realized in theutterance are removed. Strube (1998) claims that the incremental generation and pro-cessing of the S-list enables his system to handle inter- and intrasentential anaphorawithout any further specifications.

Although the S-list has the merit of avoiding ambiguities caused by the way theCb and the centering transitions interact, it is not capable of handling intrasententialanaphora without any further specifications, as claimed in Strube (1998). Stevenson etal. (2000) report experimental results pointing out cases in which focus preferences areprojected by verbs and connectives. Neither a grammatical function ordering nor aninformation-based ordering is adequate to handle such cases. To illustrate the point,we quote an example, shown in (13), from Stevenson et al. (2000). We construct the S-list ranking the elements according to grammatical function (information status wouldnot distinguish between the two proper names).11

(13) a. Keni admired Geoffj because hej won the prize.

b. Geoffj impressed Keni because hej won the prize.

In both (13a) and (13b) the pronominal resolves to Geoff, the verb argument withthe stimulus role. The ordering in the S-list in (13a), however, is Ken > Geoff, so theS-list algorithm will resolve the subsequent pronominal to the higher-ranked elementat the time of processing, in this case, Ken. In fairness to the S-list algorithm, this isa problem for any centering-based algorithm that attempts to handle intrasententialanaphora according to a fixed ranking of entities in a salience list.12

Apparently, certain discourse algorithms relying on a fixed ordering of potentialantecedents are not capable of resolving anaphora successfully. In sections 4 and 5, weargue that such cases are most commonly identified intrasententially.

2.4 Revised Algorithms for Focus Tracking/Revised Algorithms for PronounResolution

Based on previous work (Suri and McCoy 1994), Suri, McCoy, and DeCristofaro (1999)propose a methodology of extending their Revised Algorithms for Focus Tracking/Revised Algorithms for Pronoun Resolution (RAFT/RAPR) to handle focusing prop-erties of complex sentences. To determine how their framework should be extendedto handle complex sentences, they develop a methodology specifically designed to

11 This strategy was also adopted by Prasad and Strube (2000) in the implementation of the S-listalgorithm for Hindi.

12 It is conceivable that a discourse can be constructed in which the semantics will force a similar patternof resolution intersententially. Hudson-D’Zmura and Tanenhaus (1998), however, report experimentalresults that show that in such cases, sentence processing is slowed down.

327


determine how people process complex sentences. The central question they pose iswhether a complex sentence should be processed as a multiple sentence or as a singlesentence. They specifically investigated the “SX because SY” type of complex sentenceas well as its interaction with the sentences occurring in the immediately previous andsubsequent discourse.

(14) (S1) Dodge was robbed by an ex-convict the other night.

(15) (S2) The ex-convict tied him up because he wasn’t cooperating.

(16) (S3) Then he took all the money and ran.

Suri, McCoy, and DeCristofaro’s findings indicate that the pronoun resolution factswithin S2 are consistent with the expectations of both centering and RAFT/RAPR. Oncompleting the processing of the SY clause, however, the most salient entity for thefollowing discourse is not picked from SY. Based on these findings, they propose theprefer-SX hypothesis to extend RAFT/RAPR.

Although the prefer-SX hypothesis repairs the algorithm with respect to the con-struction in question, it seems to be missing a generalization regarding inconsistenciesobserved within versus across sentences. We return to this issue in Section 4.

2.5 Stevenson et al.’s Semantic/Pragmatic FocusingStevenson et al. (2000) investigate the interaction between structural, thematic, and re-lational preferences in interpreting pronouns and connectives in discourse. Stevenson,Crawley, and Kleinman (1994) have argued that the crucial factors underlying fo-cusing mechanisms in discourse are semantic/pragmatic factors. Semantic/pragmaticfocusing assumes that verbs and connectives project their own focusing preferences.Verbs project focus preferences to the entities associated with the endpoint or con-sequence of the described event. The focusing preferences of the connective dependon its meaning. For example, connectives like because direct attention to the cause ofthe previously described event, and connectives like so direct attention to the conse-quences of the event. Thus in a sentence like (17), the verb projects a focus preferencefor Bill, because Bill is the person associated with the endpoint of the event of criti-cizing. The connective, so, directs attention to the consequences, reinforcing the focuson Bill, which is then picked as the most preferred antecedent for the interpretationof the subsequent pronominal.

(17) John criticized Bill so he tried to correct the fault.

By way of demonstration, let us turn our attention to action and state verbs.The semantic/pragmatic focusing account predicts that sentences with action verbsfocus on the entity associated with the endpoint of the event, namely, the patient,independent of its structural position. This focus is maintained when the connectiveis so. In one of Stevenson et al.’s (2000) experiments, it is shown that in cases suchas (18a) the pronominal he picks the patient as its referent both when it is introducedin the previous clause as a subject and when it is introduced there as an object, asin (18b).

(18) a. Patricki was hit by Josephj so hei cried.

b. Josephj hit Patricki so hei cried.

328


A similar pattern is observed with state verbs, shown in (19), where he in the con-tinuation is interpreted as the experiencer of the event independent of its structuralposition.

(19) a. Keni admired Geoffj so hei gave him the prize.

b. Keni impressed Geoffj so hej gave him the prize.

So the experimental evidence supports Stevenson et al.’s view that the focusing prop-erties of verbs affect the interpretation of pronominals.

Hudson-D’Zmura and Tanenhaus (1998), however, report experimental resultsthat, at first blush, contradict this view. They conducted a similar experiment to testwhether subject-object or stimulus-experiencer is the crucial distinction for pronom-inal interpretation. Subjects were given sentence (20) followed by the continuations(20a)–(20b) and were asked to judge the continuations for naturalness.

(20) Max despises Ross.

a. He always gives Ross a hard time.

b. He always gives Max a hard time.

Hudson-D’Zmura and Tanenhaus’s results show that there is a strong preference forthe subject interpretation independent of the thematic role.

What are we to conclude from these inconsistent results? The results show that thesame type of verb (i.e., a state verb) in some cases projects its own focus preference(e.g., experiencer), but in other cases it does not. In order to account for the facts, oneoption would be to continue stretching structural focusing to account for the facts.Another option would be to continue stretching semantic focusing. In the followingsection, we propose an aposynthetic model for anaphora resolution in which we dividethe labor of anaphoric interpretation between the two mechanisms and define thedomains of their applicability.

3. The Proposal: Aposynthesis

3.1 Outline of the Discourse ModelWe assume that the discourse is organized hierarchically in linear and embeddedsegments as specified in Grosz and Sidner (1986). We also adopt the centering viewof local-discourse coherence to model topic continuity in discourse. According to thecentering model each segment consists of a sequence of utterances. The size of an utter-ance, however, is left unspecified. Because transitions are computed for each utterance,we will rename utterances as centering update units and argue that a centering updateunit consists of a matrix clause and all the dependent clauses associated with it. Foreach update unit a list of forward-looking centers is constructed and ranked accordingto the salience of each. Consistent with the proposed definition of unit, we argue thatentities evoked in subordinate clauses are of lower salience than entities evoked in thematrix clause and are ranked accordingly. The proposed centering specifications havethe following corollaries:

1. The linear order of subordinate clauses relative to the matrix clause doesnot affect the salience status of the entities.

329


2. Entities evoked in subordinate clauses are available as potential linksbetween the current and previous or subsequent discourse.

3. Topic shifts must be established in matrix clauses.

4. Backward anaphora in subordinate clauses is no longer “backward,” asanaphors in subordinate clauses are processed before main clausesindependent of their linear order.

Finally, we assume that anaphora across units obeys centering’s pronoun rule. How-ever, we do not adopt the BFP algorithm for anaphora resolution across units. Instead,as suggested in Section 2.1, the preferred antecedent for a pronominal in Ui is thehighest-ranked entity in Ui−1 modulo agreement features.

The remainder of this section is organized as follows. First, we briefly reviewKameyama’s tensed adjunct hypothesis, which states that subordinate clauses are in-dependent processing units and argue that on the basis of new empirical evidence thehypothesis cannot be maintained. Next, we present evidence in support of a new def-inition of the update unit. Data from English, Greek, and Japanese show that treatingsubordinate clauses as independent units yields counterintuitive centering transitionsand violations of the pronoun rule.

3.1.1 The Centering Update Unit. Defining the update unit within the frameworkof the centering model became central in very early work, because centering wasadopted and modified mainly to account for anaphora resolution. Given that anaphoricelements occur in all types of clauses, it was crucial that the size of the update unitbe constrained to enable the handling of intrasentential anaphora. To a large extent,efforts to identify the appropriate size for the unit were often dictated by needs specificto anaphora resolution algorithms.

Centering was not originally formulated, however, as a model of anaphora resolu-tion. For purposes of testing the suitability of the relevant unit in centering, it wouldbe desirable to derive a model that yields transitions that reflect our intuitions aboutperceived discourse coherence, as well as the degree of the processing load requiredby the hearer/reader at any given time in discourse processing. Reflecting degrees ofcontinuity is not a concern for anaphora resolution algorithms.

Kameyama (1993, 1998) was concerned with the problem of intrasentential cen-tering and, in particular, the definition of the appropriate update unit when complexsentences are processed. Kameyama suggested breaking up complex sentences accord-ing to the following hypotheses:

1. Conjoined and adjoined tensed clauses form independent units.

2. Tenseless subordinate clauses, report complements, and relative clausesbelong to the update unit containing the matrix clause.

With regard to her tensed adjunct hypothesis, which treated tensed adjunct clauses (forreasons of convenience, we will henceforth use the term “subordinate” to refer to thisclass of clauses) as independent units, Kameyama brings in support from backwardanaphora. She argues that the tensed adjunct hypothesis predicts that the pronoun inthe fronted subordinate clause in (21c), for example, is anaphorically dependent onan entity already introduced in the immediate discourse and not on the subject of the

330


main clause to which it is attached:

(21) a. Kerni began reading a lot about the history and philosophy ofCommunism

b. but never 0i felt there was anything he as an individual could doabout it.

c. When hei attended the Christina Anti-Communist Crusade school hereabout six months ago

d. Jimi became convinced that an individual can do somethingconstructive in the ideological battle

e. and 0i set out to do it.

This view of backward anaphora, in fact, was strongly professed by Kuno (1972),who asserted that there was no genuine backward anaphora: the referent of an ap-parent cataphoric pronoun must appear in the previous discourse. Kameyama’s (alsoKuno’s) argument is weak in two respects. First, it is not empirically tested that incases of backward anaphora the antecedent is found in the immediate discourse.Carden (1982) and van Hoek (1997) provide empirical evidence of pronouns thatare the first mention of their referent in discourse. Most recently, Tanaka (2000) re-ported that in the cataphora data retrieved from the Anaphoric Treebank, out of133 total occurrences of personal pronouns encoded as “cataphoric,” 47 (35.3%) were“first mentioned.” Among the 47 cases of first-mention cataphora, 6 instances werediscourse initial.13

Secondly, this account leaves the use of a full NP in Kameyama’s main clause(21d) unexplained (Kern and Jim have the same referent). Full NPs and proper namesoccurring in Continue transitions have been observed to signify a segment bound-ary. Assuming that segment boundaries do not occur between a main clause anda subordinate clause associated with it, the use of a full NP in (21d) remainspuzzling.

Empirical evidence in support of Kameyama’s hypothesis that tensed subordinateclauses should be treated as independent processing units was brought forth by DiEugenio (1990, 1998), who carried out centering studies in Italian. Di Eugenio (1990)proposed that the alternation of null and overt pronominal subjects in Italian could beexplained in terms of centering transitions. Typically, a null subject signals a Continueand a strong pronoun a Retain or a Shift.14

Following Kameyama (1993), Di Eugenio treats subordinate clauses as indepen-dent update units. Her motivation for doing so comes from the following example,in which the use of a strong pronoun in the main clause cannot be explained if thepreceding adjunct is not treated as an independent update unit. The translation (takenfrom Di Eugenio [1998]) is literal but not word for word. For the utterance preceding(22) the Cb(Ui−1 ) = vicinaj (neighbor-fem) and Cf(Ui−1 ) = vicinaj.

13 The Anaphoric Treebank is a corpus of a collection of news reports, annotated with, among otherthings, type of anaphoric relations. It was developed by UCREL (Unit for Computer Research on theEnglish Language) at Lancaster University, collaborating with the IBM T.J. Watson Research Center.

14 Di Eugenio collapsed the distinction between Smooth and Rough Shifts. The reader is referred,however, to Miltsakaki and Kukich (2000a, 2000b) for a discussion of the significance of Rough Shifts inthe evaluation of text coherence.

331


(22) a. Prima che i pigronii siano seduti a tavola a far colazione,Before the lazy onesi sit down to have breakfast,

b. leij e via col suoj calessino alle altre cascine della tenuta.shej has left with herj buggy for the other farmhouses on theproperty.

In Miltsakaki (2001), we report the results of a centering study in Greek. Oneof the surprising findings in this study was that a few strong pronouns appearedin Continue transitions. The result was surprising because the overall distribution ofnominal and pronominal forms revealed that weak pronouns were most common inContinue transitions, whereas strong pronouns, full noun phrases, and proper nounswere associated with Rough Shift transitions. On closer inspection, we observed that insix out of the eight instances of strong pronouns in Continue transitions, the referent ofthe strong pronoun was contrasted on some property with some other entity belongingto a previously evoked set of entities.15 Although the sample is too small to draw anydefinitive conclusions, we can at least entertain the hypothesis that strong pronounsin Italian serve a similar function. If this is true, then an alternative explanation isavailable for Di Eugenio’s data: in (22b), she, the most salient entity in the currentdiscourse, is contrasted with the lazy ones, in (22a), on the property of “laziness.” Itturns out that the hypothesis that the strong pronoun does not signal a Rough Shifttransition is confirmed by the preceding discourse, in which the “vicina” appears as themost salient entity, realized with multiple dropped subjects. The discourse immediatelypreceding (22) is shown in (23).16

(23) a. NULLj e’ una donna non solo graziosa ma anche energica e dotata dispirito pratico;and not only is shej pretty but also energetic and endowed with apragmatic spirit;

b. NULLj e la combinazione di tutto cio’ e’, a dir poco, efficace.and the combination of all these qualities is effective, to say the least.

c. NULLj si alza all’alba per sovrintendere a che si dia da mangiare allebestie, si faccia il burro, si mandi via il latte che deve essere venduto;una quantita’ di cose fatte mentre il piu’ della gente se la dorme dellagrossa.Shej gets up at dawn to supervise that the cows are fed, that thebutter is made, that the milk to be sold is sent away; a lot of thingsdone while most people sleep soundly.

We now turn to English and Greek to show that treating subordinate clauses asindependent centering units yields counterintuitive topic transitions. First, considerthe constructed example from English shown in (24).

15 We ignored one further instance of a strong pronominal in a Continue transition, as in that case thestrong pronominal headed a relative clause and its use was forced by the grammar.

16 Many thanks to Barbara Di Eugenio (personal communication) for providing me with the extra data in(23). I presume that Di Eugenio’s coding of the null realization in (23b) is based on the inferableinformation that the noun phrase ’la combinazione di tutto cio’ refers to herj qualities.

332


(24) Sequence:main-subordinate-main

a. John had a terrible headache.Cb = ?Cf = John > headacheTransition = none

b. When the meeting was over,Cb = noneCf = meetingTransition = Rough Shift

c. he rushed to the pharmacystore.Cb = noneCf = JohnTransition = Rough Shift

(25) Sequence:main-main-subordinate

a. John had a terrible headache.Cb = ?Cf = John > headacheTransition = none

b. He rushed to the pharmacystoreCb = JohnCf = John > pharmacy storeTransition = Continue

c. when the meeting was over.Cb = noneCf = meetingTransition = Rough Shift

Allowing the subordinate clause to function as a single update unit yields a sequenceof two Rough Shifts, which is diagnostic of a highly discontinuous discourse. Fur-ther, if indeed there are two Rough Shift transitions in this discourse, the use of thepronominal in the third unit is puzzling. A sequence of two Rough Shift transitionsin this short discourse is counterintuitive and unexpected given that of all centeringtransitions, Rough Shifts in particular have been shown to (a) disfavor pronominalreference (Walker, Iida, and Cote 1994; Di Eugenio 1998; and Miltsakaki 1999, amongothers), (b) be rare in corpora, to the extent that the transition has been ignored bysome researchers (Di Eugenio 1998 and Hurewitz 1998, among others), and (c) be re-liable measures of low coherence in student essays (Miltsakaki and Kukich 2000a). Inaddition, simply reversing the order of the clauses, shown in (25), causes an unex-pected improvement, with one Rough Shift transition being replaced with a Continue.Assuming that the two discourses demonstrate a similar degree of continuity in thetopic structure (they are both about “John”), we would expect the transitions to reflectthis similarity when, in fact, they do not.

Presumably, the introduction of a new discourse entity, “meeting,” in the time-clause does not interfere with discourse continuity, nor does it project a preferencefor a shift of topic, as the Cp normally does when it instantiates an entity differentfrom the current Cb. Notice that if we process the subordinate clause in the same unitas the relevant main clause, we compute a Continue transition independent of thelinear position of the subordinate clause, as the entities introduced in the main clauserank higher than the entities introduced in the subordinate clause. The computationis shown in (26).

(26) a. John had a terrible headache.Cb = ?Cf = John > headacheTransition = none

b. When the meeting was over, he rushed to the pharmacy store.Cb = JohnCf = John > pharmacy store > meetingTransition = Continue

333


Similar examples were identified in data collected from a short story in Greek(Miltsakaki 2001, 1999). Example (27), shown below, is indicative.

(27) a. Ki epeza me tis bukles mu.and I-was-playing with the curls myAnd I was playing with my hair.Cb = I, Cp = I, Tr = Continue

b. Eno ekini pethenan apo to krio,while they were-dying from the coldWhile they were dying from the cold,Cb = none, Cp = THEY, Tr = Rough Shift

c. ego voltariza stin paralia,I was-strolling on-the beachI was strolling on the beach,Cb = NONE, Cp = I, Tr = Rough Shift

d. ki i eforia pu esthanomun den ihe to teri tisand the euphoria that I-was feeling not have the partner itsand the euphoria that I was feeling was unequaled.Cb = I, Cp = EUPHORIA, Tr = Rough Shift

Again, processing the while clause in (27b) as an independent unit yields three RoughShift transitions in the subsequent discourse, reflecting a highly discontinuous dis-course. When (27b) and (27c) are processed as a single unit, the resulting sequence oftransitions for the entire discourse is a much improved Continue-Continue-Retain.

Further evidence in support of the proposed definition of the update unit comesfrom cross-linguistic observations on anaphora resolution. The most striking examplescome from Japanese.17 In Japanese, topics and subjects are lexically marked (wa andga, respectively), and null subjects are allowed. Note that subordinate clauses mustprecede the main clause. Consider the Japanese discourse (28). Crucially, the referentof the null subject in the second main clause resolves to the topic-marked subject of thefirst main clause, ignoring the subject-marked subject of the intermediate subordinateclause.

(28) a. Taroo wa tyotto okotteiru youdesuTaroo TOP a-little upset lookTaroo looks a little upset.

b. Jiroo ga rippana osiro o tukutteiru nodeJiroo SUB great castle OBJ is-making becauseSince Jiroo is making a great castle,

c. ZERO urayamasiino desuZERO jealous is(he-Taroo) is jealous.

17 Thanks to Kimiko Nakanishi for providing me with the data. In a centering study she conducted inJapanese, she also concluded that treating subordinate clauses as independent units would yield ahighly incoherent Japanese discourse.

334


In section 2.4, a similar case was also identified in English. It is repeated here as(29d). Again, the referent of he in (29d) is cospecified with ex-convict, the subject of theprevious main clause. If the because clause were processed independently, then themost salient referent available for the interpretation of the anaphoric in (29d) shouldbe Dodge. Manipulating the semantics in the second main clause to make resolution toDodge the most plausible choice does not seem sufficient to warrant felicitous pronomi-nalization, as has been shown experimentally in Suri, McCoy, and DeCristofaro (1999),demonstrated here in (30). In (30), he is not the preferred form for reference to Dodgedespite the fact that Dodge is the most natural referent for the argument of the predicatescreaming for help in this context.

(29) a. Dodge was robbed by an ex-convict.

b. The ex-convict tied him up

c. because he wasn’t cooperating.

d. Then he took all the money and ran.

(30) a. Dodge was robbed by an ex-convict the other night.

b. The ex-convict tied him up because he wasn’t cooperating.

c. #Then he started screaming for help.

The low salience of subordinate clause entities is further confirmed in the experi-mental results reported in Suri, McCoy, and DeCristofaro (1999). In their experiment,participants judge that a natural way to refer to Dodge in (31c) is by name repetition.

(31) a. Dodge was robbed by an ex-convict the other night.

b. The ex-convict tied him up because he wasn’t cooperating.

c. Then Dodge started screaming for help.

Finally, defining the main clause and its associated subordinate clauses as a singleunit points to interesting new directions in understanding backward anaphora. Withthe exception of a few modal contexts shown in (34),18 backward anaphora is mostcommonly found in preposed subordinate clauses (32) and not in sequences of mainclauses (33). From the unit definition we propose, it follows that surface backwardanaphora is no longer “backward” once the Cf list is constructed and ranked. Thereferent of the pronoun in such cases appears lower in the Cf list ranking and, infact, looks backward for an antecedent, as any other normal pronoun would. To illus-trate the point, the Cf list for (32) contains John > shower > he-referent. The pronounlooks back for an antecedent, intrasententially, and resolves to the only compatibleantecedent available, John.

(32) As soon as he arrived, John jumped into the shower.

(33) #He arrived and John jumped into the shower.

(34) He-i couldn’t have imagined it at the time but John Smith-i turned out tobe elected President in less than three years.

18 Thanks to Ellen Prince for pointing out this example.

335


3.1.2 Discourse Salience versus Information Structure. In the previous section, weargued that the linear position of the subordinate clause does not affect topic continuity.This position leads to another question: if the linear position of subordinate clausesdoes not improve topic continuity, then what is the function of clause order variation?

Let us briefly turn our attention to surface word order within a single clause. It iscommonly assumed that for each language there is an underlying canonical order ofthe basic constituents. In an SVO language like Greek, the canonical order of the verband its arguments is subject-verb-object. This, of course, is not always the attestedsurface order. In syntactic theories, it is commonly assumed that surface word orderis derived by various movement operations. Some movement operations are dictatedby the syntax of each language and are necessary to yield grammatical sentences. Itis also common, however, especially in free-word-order languages, that movementis syntactically optional and the surface word order is used to satisfy information-packaging needs (for example, to arrange the information into old-new or ground-focus or mark open propositions). Note that when this happens, it is only the surfaceword order that is altered and not the basic relation of the arguments to the predicate.To give an example from English, in (35) the internal argument of the verb (the object)has been fronted, but its original relation to the verb has remained the same.

(35) Chocolate Mary hates.

Moving to the sentential level, we entertain the hypothesis that the same principledictates the position of the clauses relative to each other. Each dependent clause standsin a specific relation to the main clause, and this relation is not altered by the orderin which the clause appears on the surface. In discourse grammars, this insight iscaptured in the discourse Lexicalized Tree Adjoining Grammar (LTAG) treatment ofsubordinate conjunctions. In discourse LTAGs, subordinate conjunctions are treatedas predicates, anchoring initial trees containing the main and the subordinate clauseas arguments. Each subordinate conjunction may anchor a family of trees to reflectvariations in the surface order of the substituted argument clauses, but the predicate-argument relation remains the same (Webber and Joshi 1998; Webber et al., 1999a,1999b).

The above discussion relates to the definition of the centering update unit in thefollowing way. The centering model keeps track of center continuations and centershifts. In other words, it keeps track of discourse salience. If we dissociate saliencefrom information structure, the relevant unit for computing salience is at the sentencelevel, which we can visualize as a horizontal level (see Figure 1). The relative order ofindependent/dependent clauses is determined by information structuring, a processpossibly orthogonal to the computing of salience. Subordinate links are not relevantto the salience mechanism. Salience is computed paratactically.

A natural consequence of this model is that one can introduce referents on thevertical level without affecting the status of the salient entity on the horizontal level.It follows that changes of topic must be established at the horizontal level. Such aconception of the salience structure suggests that text processing is not strictly incre-mental, as commonly assumed. Although it is possible that the Cf list is constructedincrementally, the final ranking is determined only after the sentence is complete.

Admittedly, the distinction between discourse salience and information packagingis hard to establish because of the inevitable overlap between information status andsalience: attention centers, for example, tend to be discourse old. Still, there are otheraspects of information packaging pertaining to clause order (e.g., temporal or logicalsequences, open proposition frames inherited from previous discourse) that do not

336


Figure 1Salience model.

necessarily relate to the salience of the participating entities. Although a great dealof additional work is required to understand the precise nature of the interactionbetween salience and information structure, we believe that we obtain a significantgain in keeping the two processes distinct.

3.2 Outline of the Anaphora Resolution ModelIn Sections 2 and 3, we discussed a number of challenging cases for anaphora res-olution, including some puzzling experimental data. We raised the question of howthe data are to be reconciled. We are now able to offer an explanation. The basicidea is that topic continuity and intrasentential anaphora are handled by two distinctmechanisms. Topic continuity is computed across centering update units. Anaphoricreference spanning across update units relates to topic continuity and is therefore de-termined structurally in accordance with centering rules and constraints. Within theunit, anaphora is constrained by focusing preferences projected by the matrix predi-cate and the extended arguments of the predicate that can be locally realized throughsubordination.

This basic outline is sufficient to explain (most of) the data we have seen so far. Theexperiments reported in Stevenson et al. (2000), which show a main effect of thematicfocusing, involve the interpretation of anaphoric expressions in subordinate clauses.On the other hand, Hudson-D’Zmura and Tanenhaus’s (1998) experiments on similartypes of verbs show a main effect of structural focusing. The difference between thetwo sets of experiments is that Hudson-D’Zmura and Tanenhaus’s (1998) experimentsinvolve sequences of main clauses, whereas in Stevenson et al. (2000) the relevantexperiments involve subordinate clauses. Furthermore, Stevenson et al. (2000) reportresults on a different set of experiments showing a main effect of structural focusing,and these are precisely the experiments containing sequences of main clauses. Further,Suri, McCoy, and DeCristofaro’s (1999) “SX because SY” construction indicates thatthe referent appearing in the subordinate clause is not the preferred focus in thesubsequent discourse, whereas resolution to the subject of the main SX clause yieldsthe desired interpretation.

The remainder of this section is organized as follows. First, we provide definitionsfor the basic tenets of the model we propose and describe the basic steps required forcombining the two mechanisms in a single anaphora resolution algorithm. Next, we

337


discuss some remaining issues raised by the English connective so and certain typesof preposed subordinate clauses.

3.2.1 Algorithm and Model Specifications. Discourse consists of a sequence of seg-ments. Each segment consists of a sequence of centering update units. A single center-ing update unit consists of one main clause and all its associated dependent clauses.Dependent clauses are of three types: sentential complements of verbs, relative clauses,and subordinate clauses. Sentential complements of verbs and relative clauses are iden-tified syntactically. Subordinate clauses are introduced with subordinate conjunctions.To identify subordinate conjunctions, we apply the reversibility test: a tensed clauseis introduced by a subordinator when the clause it introduces can be preposed.19

For example, in (36), although is classified as a subordinator and the although clauseis classified as a subordinate clause because placing the although clause before themain clause retains grammaticality. Conversely, however in (38) is not classified as asubordinator, because preposing the clause with which it is associated yields ungram-maticality.

(36) John traveled by air although he is afraid of flying.

(37) Although he is afraid of flying, John traveled by air.

(38) John traveled by air. However, he is afraid of flying.

(39) #However, he is afraid of flying. John traveled by air.

Update units are identified and numbered. For each identified update unit the listof forward-looking centers is constructed and its members are ranked according tothe ranking rule for English. The “M” prefix stands for “main clause,” and the “S”prefix stands for “subordinate clause.” The relevant ranking of the various types ofdependent clauses is currently left unspecified.

Ranking Rule for EnglishM-Subject > M-indirect object > M-direct object > M-other >S1-subject > S1-indirect object > S1-direct object > S1-other >S2-subject > . . .

For complex NPs, we assume left-to-right ranking of entities, as suggested in Walkerand Prince (1996).20

Given the above input for N units, Ui=1 ...N, the anaphora resolution algorithmstarts at the last identified unit. The basic steps are specified below. Some of the stepsrequire information that is obtainable by currently available natural language systems:syntactic parsers, morphological analyzers, automated proper name identification, andelectronic lexical databases such as WordNet (to check animacy, for example, as wouldbe necessary for the ranking of entities in Greek). Others, such as understanding thefocusing preferences of verbs and connectives as well as identifying thematic roleswill, of course, await further research.

19 “Reversibility” is identified as a characteristic of subordinate clauses in Quirk et al. (1972).20 Complex NPs are noun phrases containing multiple nouns, for example, John’s father.

338


0. Start at the last identified unit Ui with i = N.1. Identify pronominal expressions in the rightmost subordinate

clause.2. Input antecedents from the Cf list.3. Apply grammar-driven constraints (number and gender

agreement, contra-indexing, etc.) to reduce list of potentialantecedents.

4. Resolve from right to left to the first available antecedent insidethe subordinate clause. Output unresolved pronominals.

5. Using the Cf list resolve pronominals according to semantic fo-cusing constraints. Output unresolved pronominals.

6. If there is another subordinate clause to process go to step 1.7. Identify pronominals in the main clause. Apply grammar-driven

constraints (number and gender agreement, contra-indexing,etc.) to reduce list of potential antecedents. Resolve from rightto left to the first available antecedent inside the current clause.Output unresolved pronominals.

8. Input Cf list of potential antecedents from previous unit.9. Apply grammar-driven constraints to reduce list of potential

antecedents.10. Resolve pronominals starting from the leftmost to the highest-

ranked element of the list of available antecedents.11. If an antecedent is found go to step 13.12. If the list of potential antecedents is empty and there is a unit

to process go to step 8, else mark unknown.13. If Ui is the first unit U1 terminate, else start processing Ui−1

and go to step 1.

By way of demonstration, we apply the algorithm to resolve the anaphoric ex-pressions in discourse (2), repeated here in (40)–(42).

(40) Dodge was robbed by an ex-convict.

(41) The ex-convict tied him-3 up because he-2 wasn’t cooperating.

(42) Then he-1 took all the money and ran.

• Step 0 applies. Move to step 1.

• No subordinate clause is identified. Jump to step 7.

• Step 7 applies. The pronoun HE-1 is identified. There is no potentialantecedent in the current clause. Move to step 8.

• Step 8 applies. The Cf list from the previous unit containsEX-CONVICT > DODGE.

• Step 9 applies. Grammar constraints do not reduce the list of potentialantecedents.

• Step 10 applies. HE-1 resolves to the EX-CONVICT.

• Step 13 applies. Move to step 1.

• Step 1 applies. The pronoun HE-2 is identified.

339


• Step 2 applies. The Cf list is empty (it contains only the unresolvedpronoun HE-2).

• Steps 3 and 4 apply vacuously. There are no potential antecedents in thecurrent clause.

• Step 5 applies. The Cf list contains HIM-3 > EX-CONVICT because ofsemantic focusing. HE-2 resolves to HIM-3.

• Step 7 applies. HIM-3 is identified. Grammar constraints apply andcontra-index EX-CONVICT with HIM-3.

• Step 8 applies. The Cf list from the previous unit containsDODGE > EX-CONVICT.

• Step 9 applies. Grammar constraints do not reduce the list of potentialantecedents.

• Step 10 applies. HIM-3 resolves to DODGE.

• Steps 11–13 apply. The algorithm terminates.

3.3 Comparison with Related AlgorithmsThe crucial difference between our approach and related anaphora resolution algo-rithms is in the treatment of subordinate clauses. Whereas steps 7–10 are similar toother approaches that opt to resolve a pronoun to the highest-ranked element of theCf list of the previous clause, the resolution process described in steps 0–7 and theCf ranking assumptions described earlier are not. As indicated in the ranking rulefor English set forth in Section 3.2.1, (a) subordinate clauses are part of the same unitcontaining the main clause with which they are associated, and (b) there is a single Cfranking list for both the main and the subordinate clauses. Because the entities in thesubordinate clauses rank lower than the entities in the main clause, the linear positionof the subordinate clause does not affect the resolution process. We have seen that this“restoring” of a basic clause order results in virtually eliminating backward anaphora,which in other approaches requires special treatment.21 Also, intrasentential anaphorais preferred in the cases of anaphoric elements occurring in subordinate clauses butnot in main clauses (assuming grammatical filtering), again irrespective of their linearorder.

We will now demonstrate these differences with respect to Lappin and Leass’s(1994) and Hobbs’s (1978) algorithms, which are conceptually the closest to our ap-proach. Lappin and Leass’s RAP (Resolution of Anaphora Procedure) algorithm ap-plies to the output of McCord’s (1990) Slot Grammar parser and utilizes measures ofsalience derived from syntactic structure and a simple model of attentional state. Po-tential anaphor antecedents receive a salience score on which they are evaluated. Thescoring system penalizes backward anaphora and rewards parallel syntactic positionsand intrasentential antecedents (sentence recency).

As we have already mentioned, backward anaphora need not receive any specialtreatment in our approach. Lappin and Leass penalize cases of backward anaphoraseverely, which seems to work well on empirical grounds, presumably because back-ward anaphora is rather rare. In absence of an explicit method of identifying realcases of backward anaphora, however, the system is likely to miss such cases. In our

21 Assuming that backward anaphora is restricted to subordinate clauses. Special treatment is required forthe but clauses discussed in section 3.1.1, example (34).

340


approach, this is not a problem, because the Cf ranking of the processing unit im-plicitly identifies all real cases of backward anaphora and converts them into forwardanaphora.

Further, some of the limitations of the system discussed by Lappin and Leassinvolve cases of intersentential anaphora such as the following:

(43) a. This green indicator is lit when the controller is on.

b. It shows that the DC power supply voltages are at the correct levels.

The RAP algorithm resolves the pronoun it in (43b) to the controller in (43a). This isbecause, in RAP, the subject of the main as well as the when clause in (43a) are of equalsalience. In this case, the controller wins because it is more recent. In our approach, itwould resolve to the highest-ranked entity of the previous unit, which in this case iscorrectly identified as the green indicator. This is because the when clause is not treatedas an independent unit. The entities evoked in the when clause are linearly but notstructurally more recent.

Hobbs’s (1978) syntactic algorithm is based on a well-defined search procedure(left-to-right in most cases, breadth-first) applied on the surface parse tree. The al-gorithm has three main components. The first component treats reflexive pronounsby constraining the search procedure with special configurational requirements. Thesecond component takes over when the antecedent of an anaphor is to be found inprevious sentences, and the third component searches subparts of the parse tree incycles until the highest clause is reached.

Intersententially, Hobbs’s syntactic algorithm favors subjects over objects, as sub-jects are higher up in the parse tree than objects. Intersententially, our approach andHobbs’s algorithm would opt for the same type of antecedent. As Lappin and Le-ass (1994) have pointed out, however, the syntactic search procedure seems to workpretty well in English because grammatical order corresponds to phrase order. Forother languages, either free-word-order languages like Greek or languages in whichsalience is determined by other factors (e.g., information status, as has been argued forGerman [Strube 1998]), Hobbs’s search procedure would fail, because it is too rigid toaccommodate linguistic variation in marking salience. Even for languages like English,the relevant salience of entities may be undermined by nonsyntactic factors. As hasalready been suggested by Turan (1998), among others, certain types of NPs are lesssalient than others independent of their grammatical function (e.g., indefinite quan-tified expressions and impersonal pronouns). The flexibility of constructing lists ofentities according to salience both optimizes the capabilities of an anaphora resolutionalgorithm and is best suited to accommodate the multiplicity of factors that may haveto be taken into account in determining reference salience.

Hobbs’s algorithm is, in effect, similar to our approach in the treatment of subor-dinate clauses. Subordinate clauses belong to the same parse tree as the main clause towhich they are subordinate. This is equivalent to our claim that subordinate clauses arenot independent processing units. With respect to backward anaphora, in particular,Hobbs’s use of the “command” relation achieves the same result as our lower rank-ing of entities appearing in subordinate clauses. The subject of a subordinate clausewould be lower in the parse tree than the subject of the main clause, independentof the linear position of either. So, for example, in (44), the pronoun would correctlyresolve to Susan. In a case like (45), however, Hobbs’s algorithm would always resolvethe pronoun to Susan, since the search procedure has no way of making a distinc-tion between different types of subordinate connectives (or verbs) and their effect onreference salience.

341


(44) After she phoned Barbara, Susan went out for dinner.

(45) Susan criticized Barbara because she was lazy.

3.4 Some Remaining IssuesAs mentioned above, the proposed model for anaphora resolution accounts comfort-ably for the results reported in Stevenson et al. (2000) except, however, for the exper-iment involving the connective so.

According to the reversibility test, so is classified as a subordinate conjunctiondepending on its interpretation. In English, so denotes two relations: consequence andpurpose. The examples below indicate that only purpose-so behaves as a subordinateconjunction. 22

(46) I had to give up my job so I could be happy again.

(47) So I could be happy again, I had to give up my job.

(48) I had just been to the bank, so I had money.

(49) #So I had money, I had just been to the bank.

The anaphora resolution model we propose predicts that the interpretation ofpronouns in consequence-so sentences is determined structurally. This prediction wasnot borne out. Stevenson et al. (2000) report a main effect of semantic focusing inconsequence-so continuations.

There are two options available to explain the data. First, we may hypothesize thatsubordination is determined on structural grounds, in which case it is likely that lan-guages may arbitrarily characterize their set of subordinate conjunctions. Under thisoption, we may hypothesize that so in English is uniformly a subordinate conjunctionand then set out to investigate the implications of such a hypothesis on empiricalgrounds. Alternatively, we may hypothesize that the crucial factor in characterizingsubordination is by its semantic properties, that is, the type of relation it establisheswith the proposition denoted in the main clause. This second option seems intuitivelyappealing and more promising in explaining this otherwise puzzling linguistic phe-nomenon, namely, the structural distinction between main and subordinate clauses. Itruns into the following problem, however.

In Modern Greek, the equivalent conjunction for the English so is etsi or ki etsi (=‘and so’), which is not polysemous and is not a subordinate conjunction. Greek etsilinks clauses paratactically (i.e., links sequences of main clauses). The examples belowshow that Greek behaves differently from English in the so cases.

(50) #I Mariai htipise tin Elenij, ki etsi NULLj evale ta klamata.the Maria hit the Eleni and so she put the tears.Mariai hit Elenij and so shej started crying.

(51) I Mariai xilokopithike apo tin Elenij ki etsi NULLi evale ta klamata.the Maria was-hit by the Eleni and so she put the tears.Mariai was hit by Elenij and so shei started crying.

22 Although in many cases preposing a purpose-so clause seems unnatural, at least for some nativespeakers of English preposing of a purpose-so clause is possible, given the appropriate context. For allnative speakers we have consulted, there is a marked difference in the acceptability of (47) and (49).

342


The Modern Greek data show that the null subject in the so clause cannot be inter-preted as the object of the previous clause. If subordination was to be defined onsemantic grounds, then we should not expect focusing differences between the twolanguages, but in fact we do.

Finally, we notice that Greek is much like English when the second clause is linkedthrough other types of subordination, as shown in (52)–(53):

(52) I Mariai htipise tin Elenij giati NULLj ekane ataxies.the Maria hit the Eleni because she did naughty-thingsMariai hit Elenij because shej was being naughty

(53) I Elenij xilokopithike apo ti Mariai giati NULLj ekane ataxies.the Eleni was-hit by the Maria because she did naughty-things.Elenij was hit by Mariai because shej was being naughty.

The reason for the difference between the two languages with respect to so clauses ishard to explain. This difficulty in understanding the cross-linguistic variation is alsotelling of our fundamental lack of understanding subordination in languages. In thisarticle, we do not claim to understand the intricacies of subordination any better. Inthe next section, it is shown that the distinction between main and subordinate clausesis in the right direction. It is not yet clear, however, what property of subordination—structural, semantic or other—is responsible for the pattern we observe. We will leavethis issue open for future work.

Another issue that requires special attention in the proposed account pertains tosome special cases of preposed subordinate clauses. Example (54) presents a problemfor the proposed model because the antecedent of the subject pronoun in the matrixclause is the subject of the preposed subordinate clause.

(54) After Susan phoned Barbara, she went out for dinner.

The ranking in the Cf list for (54) is she-referent > Susan > Barbara. In effect, whatwe are faced with here is analogous to backward anaphora. In its current form, how-ever, the proposed algorithm would process the subordinate clause first and wouldthen move to the matrix clause. The matrix clause contains a pronoun and no pos-sible antecedent, so on completing the processing of the unit, the algorithm wouldoutput the unresolved pronoun from the matrix clause and would continue searchingfor an antecedent in the previous unit. Such cases can be identified easily by evenshallow parsing and be fixed locally by forcing resolution to the highest entity in thecurrent unit (i.e., Susan). Also, as a reviewer has suggested, the algorithm presentedin Section 3.2.1 could be modified so that in step 2 the Cf list includes all possibleantecedents from the current utterance Ui. With this modification, (54) would be pro-cessed correctly, but as the same reviewer points out, this does not explain the contrastin (55):

(55) a. Susan phoned Barbara. Then, she went out for dinner.

b. Susan phoned Barbara before she went out for dinner.

c. After Susan phoned Barbara, she went out for dinner.

Example (55a) is an instance of intersentential anaphora, and there is a subjectreference for the pronoun as predicted. Example (55b) is a case of intrasentential

343


anaphora, and there is no clear subject reference. Example (55c) is another instanceof intrasentential anaphora, but in this case the subject preference is clearly on a parwith the intersentential case in (55a). Whatever required modification to the algorithmwill prove to be more useful, the fact remains that the similarity between (55a) and(55c) remains unexplained in purely structural terms. We suspect that the differencebetween (55b) and (55c) and the similarity between (55a) and (55c) is the result ofan interaction with a discourse function of subordinate clauses. Subordinate clausesnormally convey background information and do not by themselves move the nar-rative forward. They also have the property of enabling information to appear in a“nonnatural order” with respect to the event(s) of the main clause. A “natural order”for temporal connections would be to express events in the order in which they hap-pened. For causal connections, a natural order would be to express the cause beforethe effect. So it seems plausible to hypothesize that subordinate structures can be usedto introduce background (or presupposed) information and even discourse-new char-acters without disturbing the narrative structure of the discourse and the salience ofthe centers of attention already established in the narrative. If this line of thinkingis on the right track, then it is possible that the similarity between (55a) and (55c) isdue to the fact that both sequences of clauses reflect the linear succession of events.The preposed after-clause does not disturb the natural temporal order of events bothof which are predicated of the same center, which in this case is introduced in thesubordinate clause. Further empirical work is clearly needed to evaluate this line ofexplanation.

4. Empirical Studies

In this section we report the results of two empirical studies designed specificallyto test the central hypotheses of the proposed model: (a) that anaphoric referencethat spans across centering update units is determined structurally (as specified inSection 3.2.1) and (b) that subordinate clauses do not form independent processingunits. In the first study, in Section 4.1, we report the results of a sentence completionexperiment in English. We quantify over the interpretation of a subject pronoun toone of two ambiguous antecedents evoked in the preceding clause. Two factors wereanalyzed, type of clause and semantic type of connection, in four conditions. The re-sults show a strong main effect for type of clause. In the main-clause conditions, thesubject pronoun resolved to the subject antecedent of the previous sentence. In thesubordinate-clause conditions, the interpretation of the subject pronoun was varied.In the second study, we tested the same hypotheses on a Greek corpus. We selectedthree types of subordinate clauses under conditions similar to the ones set in the ex-perimental study. We contrasted anaphoric interpretation in subordinate clauses withanaphoric interpretation in main clauses. The results provide strong evidence for theaccuracy of the centering-based algorithm proposed for anaphora resolution acrossunits. The results also confirm that anaphoric interpretation in subordinate clauses isnot determined structurally.

4.1 Experimental DataThe aim of the experiment discussed in this section is to investigate the hypothe-sis that subject pronouns in main clauses follow a different pattern of interpretationfrom subject pronouns in subordinate clauses. The participants in this experiment readsentences containing sequences of two clauses. In two of the four conditions of theexperiment, the sequence consisted of one main and one subordinate clause. In theother two, the sequence consisted of two main clauses. In all the conditions, the first

344


clause introduced two individuals of the same gender. The second clause containeda subject pronoun of the same gender as the individuals introduced in the first mainclause. Participants were asked to complete the second clause. We were interested inthe interpretation of the subject pronominal in the second clause.

For this experiment, two subordinate conjunctions were selected from two se-mantic classes, namely, time and contrast. Also, two adverbial conjunctions, then andhowever, were selected from the same semantic classes to introduce the second clausein the main-main conditions.

We predicted that the pronominal in the main clause would consistently be inter-preted as the subject of the preceding main clause. We also predicted that the interpre-tation of the subject pronoun in the subordinate clause would vary. In all critical itemsthe verb of the main clause belonged to the same category (see below), so any varia-tion pattern observed across the two subordinate conjunctions would be attributed tothe focusing preferences projected by the semantics of the subordinate conjunctions.

The grouping of connectives into semantic types, namely, then, when for time andhowever, although for contrast and structural types, then, however for main clauses andwhen, although for subordinate clauses, enables us to compare and contrast the effect oftwo factors: semantic type and type of clause. A main effect of semantic type wouldmean that the interpretation of the subject pronoun is primarily determined by thesemantics of the connectives. Conversely, a main effect of type of clause would meanthat the interpretation of the subject pronominal is primarily determined by the typeof clause.

4.1.1 Study and Results. Sixteen adults, native speakers of English, participated inthe experiment. They were asked to perform a sentence completion task as describedin the previous section. Each participant received a form containing 36 fillers and 12critical items, 6 per type of clause and 6 per semantic type. The example below demon-strates a complete set of conditions; each form contained three such sets interspersedwith fillers. Each main clause appeared in all conditions across participants, and eachcondition appeared three times within participants. The order of appearance of criticalitems was varied in two LISTS (two variations in the order of appearance of criticalitems).

a. The groom hit the best man violently. However, he. . . (contrast, main-main)b. The groom hit the best man violently

although he. . . (contrast, main-subordinate)c. The groom hit the best man violently. Then, he. . . (time, main-main)d. The groom hit the best man violently when he. . . (time, main-subordinate)

All main clauses contained an action verb with two human arguments of the samegender. An adverbial phrase was added at the end of the main clause to achieve nat-uralness. The arguments of the verbs were always realized as role NPs to minimizereferential ambiguity in the continuations. In general, sentence continuation involv-ing role NPs tends to disambiguate the intended reference. Nevertheless, an averageof two ambiguous continuations were identified per participant. On completing theexperiment, the participants were asked to identify explicitly the intended referent ofthe ambiguous continuations.

Figure 2 summarizes the distribution of the data within participants. The y-axisshows the number of references to the main-clause subject. The lighter grey columnsshow the distribution of the interpretation of the main-clause subject pronoun, and the

345


Figure 2Experiment on anaphora.

darker grey columns show the distribution of the interpretation of the subordinate-clause subject. For convenience, we labeled the main-main clause condition parataxisand the main-subordinate clause condition hypotaxis.23

Analyses of variance revealed a strong main effect of type of CLAUSE: the pronounresolved to the subject of the previous clause significantly more often when it appearedin a main clause (F = 74.16, df = 1, 15, p < .0001). Semantic type showed only amarginal effect (F = 4.59, df = 1, 15, p < .049).

With regard to the comparison based on the semantic similarity between how-ever/although and then/when, it turned out that in several cases when clauses wereassigned a causal interpretation. For example, the when clause continuation in (56) isinterpreted as giving the cause of the event described in the preceding main clause,rather than the time specification.

(56) The father shook the son vigorously when he saw him lying on theground.

For this reason, it was hard to pursue the semantic-based comparison between theparatactic then and the hypotactic when, and we therefore did not make any furtheranalyses within the semantic-type factor. Since it would be hard to control the inter-pretation of the connectives used in this type of experiment, in future experimentswe plan to give up on the effort to pair connectives semantically. Instead, we plan to

23 Parataxis and hypotaxis are terms borrowed from traditional grammars to describe the two types ofconnections. Main clauses are linked to each other through parataxis, whereas a subordinate clause islinked to its superordinate clause through hypotaxis.

346


include a larger number of subordinate conjunctions in order to be able to generalizemore reliably over the whole class of subordinate conjunctions.

4.1.2 Discussion. The results of this experiment show that structural focusing is promi-nent across main clauses. The analyses of variance revealed a strong main effect forthe type of clause despite the fact that, in the experiment, we included main-clausecontinuations introduced with however. Given the verb type in the first main clause,we in fact stretched structural focusing to its limits, as it would be reasonable to expectthat the contrast relation established with however would shift attention to the secondindividual. For example:

(57) The Popei tapped the priestj on the shoulder. However, hej ignored him.

Still, the results clearly indicate that the type of connection affects the interpretationof the pronoun in the second main clause across, or despite, semantic types. Also, thevariation of pronominal reference across types of subordinate conjunctions indicatesthat the interpretation of anaphoric expressions in subordinate clauses is determinedby other, apparently nonstructural, factors.24

Finally, the results are consistent with the hypothesis that structural focusing de-termines the interpretation of pronominals across, but not within, processing units.According to the discourse model suggested in Section 3.1, a main clause is an in-dependent processing unit, so a sequence of main clauses constitutes a sequence ofunits. On the other hand, a subordinate clause does not form an independent unit, soa sequence containing a main and a subordinate clause is simply internal to the unit.The reference variation observed in the subordinate conditions is then expected fromthe model.

Unfortunately, in this experiment, we lost the contrast for the type of connectionwithin semantic types because of the noise created by the causal interpretation ofsome when clauses. To enable the conclusions of this experiment to be generalizedacross the entire classes of paratactic and hypotactic connectives, this experiment willbe repeated with a larger set of connectives.

An interesting extension of this work would be to investigate whether structuralfocusing is active within units, before it is eventually overridden by semantic focusing,or whether focusing preferences projected by the semantics of the verbs and connec-tives are immediately accessed during on-line processing.

4.2 Corpus-Based DataThe corpus-based study reported in this section is based on a Greek corpus built withtext from Greek newspapers available on the Web. The corpus consists of approxi-mately 800,000 words downloaded from the sites of the Greek newspapers Eleftherotipiaand To Vima.

The aim of the corpus study is similar to that of the experimental study reportedin the previous section. Greek allows dropped subjects, which yields higher referen-tial ambiguity than English pronouns. We were therefore able to collect a reasonablenumber of tokens fulfilling the conditions of the experimental study and to test ourhypotheses against naturally occurring data.

24 It is understood that in the main-subordinate clause sequences, the subordinate clause is linked to themain clause in question and not to some subsequent main clause (as, for example, in cases in whichthe subordinate clause is preposed).

347


As in the experimental study, we wanted to compare and contrast the interpre-tation of anaphoric expressions in a main clause with the interpretation of anaphoricexpressions in a subordinate clause. Unlike in the experimental study, however, thesearch of anaphoric expressions in main and subordinate clauses was not restricted tosubject pronouns.

4.2.1 Data Collection and Coding. Greek has two pronominal systems: weak andstrong. Weak pronominals include dropped subjects and object clitics. Clitics are markedwith case, gender, and number features and attach to the verb. Direct-object clitics arecase-marked “accusative,” and indirect or “dative” clitics are case marked “genitive.”Strong pronominals are also marked with case, gender, and number features, but syn-tactically, they behave as common nouns. Their functions are mentioned in Sections 2.2and 3.1.1. In this study, only weak pronominals were included.

We established the following requirements for the data set of this study: (a) thesubordinate clause or second main clause contains a third-person dropped subject orweak pronominal, and (b) the preceding main clause or any of its other associated sub-ordinate clauses contains at least two competing antecedents. A competing antecedentis defined as a full noun phrase, dropped subject, or weak pronominal that agrees ingender and number with the anaphoric expression.

For anaphoric reference in main-main and main-subordinate sequences, ideallywe would have liked to include only those tokens in which the second main or sub-ordinate clause under investigation was preceded by a unit containing only a mainclause. Imposing this extra constraint, however, would have invalidated a large num-ber of the already limited number of tokens, so we decided against doing so. Althoughconducting a second pass of the data, with the purpose of studying further the char-acteristics of these antecedents, would have been useful, for the purposes of this studyit was not crucial. Further, a consistent pattern of reference in main-main sequencesincluding cases in which competing antecedents are present in intervening subordi-nate clauses provides further evidence that entities introduced in subordinate clausesdo not override the salience of the main-clause entities. We will provide an exampleto illustrate the point in the next section.

For the data set with main-subordinate sequences, we extracted three types of sub-ordinate clauses introduced by the following subordinate conjunctions: otan (‘when’),yati (‘because’), and oste (‘so that’). The final data set included only tokens that fulfilledthe requirements described above.

For the data set with main-main sequences, we randomly selected files from thecorpus subdirectories and included tokens that fulfilled the requirements describedabove. The selection process was terminated when the number of qualifying tokensapproximated one hundred.

Two coders, both native speakers of Greek, marked on the data set the antecedentof the anaphoric expressions. One of the coders was the author and the other was anaive, nonlinguist speaker of Greek. Intercoder reliability was particularly high (98%).We attribute the high intercoder reliability to the fact that discourse-deictic expressions,known to lower intercoder reliability, were not included in this study.25 The few cases ofdisagreement between the coders involved either instances perceived as ambiguousby the coders or abstract complex NPs about which there was disagreement as to

25 Discourse-deictic expressions include demonstratives, such as this and that, used to refer to chunks ofprevious discourse. Discourse-deictic expressions in Greek are identical in form with neuter strongpronominals. Dropped subjects can refer to discourse-deictic expressions, but such cases were excluded.

348


whether the antecedent was the possessor or the possessee. Such cases were excludedfrom the final data set.

The final dataset included 88 instances of main-main sequences and 108 instancesof main-subordinate sequences broken up as follows: 48 otan clauses, 17 yati clauses,and 43 oste clauses.

4.2.2 Ranking Antecedents and Coding. Based on earlier work on the ranking ofentities in Greek (Miltsakaki 1999), the competing antecedents were ranked accordingto the following rule:

Ranking Rule for GreekEmpathy > Subject > Indirect Object > Direct Object >Indefinite Quantified NPs, Nonspecific Indefinites

Under Empathy were classified dative subjects of psych verbs. Such verbs are easilyidentified from a normally short exhaustive list that can be enumerated for each lan-guage. In our data, we encountered only the verb like from this verb category.26

All entities introduced in subordinate clauses associated with the main clause areranked using the same rule but lower than the main-clause entities. So, for example, ifthe evoked entities are main subject, main object, and subordinate subject, the Cf listis ranked as follows: main subject > main object > subordinate subject. It is not clearwhat the ranking would be in cases of multiple subordinate clauses, but this extraranking specification was not crucial for the current study.

What was crucial for the study was the ranking of entities evoked within complexNPs. Greek complex NPs are normally constructed with two nouns: the “possessor,”marked with genitive, and the “possessee,” marked with nominative, accusative, or,more rarely, genitive, depending on its grammatical role. The possessee always pre-cedes the possessor.27 Noun-noun modification is not allowed in Greek. In complexNPs, animate referents rank higher than inanimates. In all other cases, possessor rankshigher than possessee. For clarification we present an example below, followed by theranking of the evoked entities:

(58) I mitera tis Marias ipe sto Yani oti o Giorgos den tha erhotan.the mother of-the Maria said to-the John that the George not would comeMaria’s mother told John that George would not come.

Maria > mother > John > George

The salience ranking as specified above was then used for a second pass of codingdone by the author. For each set of candidate antecedents, the intended referent wasmarked as either “preferred antecedent” (designated by “Ap”) or “nonpreferred an-tecedent” (designated by “Anp”). The referent of an anaphoric expression was marked

26 By way of demonstration, the expression I like John in Greek is glossed as “me-genitive like-3rd singularJohn-nominative.” In the Greek example, the experiencer of the psych verb is analyzed as subjectdespite its genitive marking. Such subjects are known as dative subjects. Modern Greek has lost thedative case, whose function is now performed by the genitive case.

27 We use the terms possessor and possessee here for convenience to label the structural position of nounsin complex NPs. These terms, however, do not always describe the semantic relationship between twonouns. For example, in John’s participation, John can hardly be characterized as the possessor, but inGreek participation would always precede John and would be case-marked genitive.

349


as a preferred antecedent when it was the highest-ranked entity in the set of competingantecedents. The referent of an anaphoric expression was marked as a nonpreferred an-tecedent when it was not the highest-ranked entity in the set of competing antecedents.In most cases, the set of candidate antecedents included only two candidates, so sub-categorizing nonpreferred antecedents was not crucial.

In what follows, example (59) is demonstrative of cases in which the referent ofthe anaphoric expression in the second main clause is marked “Ap.” The competingantecedents in (59a) are ta opla and anthropus, because they have the same number andgender as the anaphoric tus in (59b). The NP ta opla is the preferred antecedent becauseit is the highest-ranked element in the list of potential antecedents and is the intendedreferent of the anaphoric. We report this particular example for the additional reasonthat it shows that, outside complex NPs, animacy is not a factor in determining theranking of entities, even in cases in which the semantics of the verb taking the referentof the anaphoric as an argument favors the human, in this case, antecedent. Example(60) also demonstrates a case of reference to Ap. Here, the competing antecedents areboth male characters and semantically plausible subjects of the verb egrafe (‘wrote’).Note that the assumed ranking receives further support with this example, since theanaphoric resolves to the subject of the previous clause and not to the most recent,equally plausible entity.

(59) a. [Ta opla]i ine kataskevasmena ya na skotonun [anthropus]j.the guns are made in-order to kill people.Gunsi are made to kill peoplej.

b. Aftos ine o skopos [tus]i.This is the goal their.This is theiri goal.

(60) a. [O Turen]i vriskete apo filosofiki apopsi ston antipoda.the Turen is-placed from philosophical view at-the opposite-side[tu Popper]j.of-the Popper.From a philosophical point of view Tourrainei is the very opposite ofPopperj.

b. Prosfata [0]i egrafe oti iparhun dio idon dianoumenirecently 0 wrote that there-are two types of-intellectuals.Recently, hei wrote that there are two types of intellectuals.

In (61) the referent of the dropped subject in (61b) is marked “Anp.” The list ofcompeting antecedents in (61a) contains PAOK, the name of a football team, and PikulinOrtith, the name of a player, both being singular and masculine. In this case, numberagreement with the verb is sufficient to create ambiguity, because Greek verbs aremarked for number but not gender. Also, in Greek, subject collective nouns markedsingular always take a singular verb. The intended referent is Anp, because it is rankedlower than the subject PAOK.

(61) a. Ya mia sira praxeon [o PAOK]i kali [ton Pikulin Ortith]j naFor a series of-deeds the PAOK summons the Pikulin Ortith to

350


apologithi amesa,confess immediately,PAOK is asking Pikulin Ortith to confess immediately for a series ofthings,

b. yati [0]j ehi prokalesi megisti agonistiki ke ithiki zimia.because 0 has caused enormous competitive and moral damage.because [he]j has caused enormous damage (to the team) both morallyand in the championship.

Finally, example (62) demonstrates that competing antecedents in dependent clausesdo not override the salience of main-clause antecedents. Note that [i kinonikes igesies]is a perfectly natural candidate for the subject of the verb pistevun.

(62) a. [I esiodoxi]i pistevun oti ehun dimiurgithi [i kinonikes igesies]j

the ambitious believe that have been-created the social leadershipspu mporun na antiparatethun stin katestimeni exusia.which can to object to-the established leadership.[The ambitious ones]i believe that there have been formed [socialauthorities]j

which can stand up to the established leadership/political power.

b. [NULL]i pistevun oti o agonas tus den ehi akrivos kerdithi allaNULL believe that the fight their not has exactly been-won butoti NULL vriskete se “dromo horis epistrofi.”that NULL is-found in “road without return.”[They]i believe that their fight has not exactly been won but that it is ata point with no return.

In the next section we present the results of the analysis of the distribution ofanaphoric references based on the values of Ap and Anp.

4.3 Results and DiscussionTable 2 shows the distribution of anaphoric reference in the experiment described inSection 4.2. The first column shows the number of times the anaphoric expressionresolves to the preferred antecedent (Ap). The second column shows the number oftimes the anaphoric expression resolves to a nonpreferred antecedent (Anp), and thethird column summarizes the total number of tokens per condition.

The corpus-based results support the hypothesis that anaphora does not obeythe same rules in main and subordinate clauses.28 Clearly, the preferred antecedentas defined structurally is a strong predictor of the referent of main-clause anaphoricexpressions, whereas the picture appears more complicated in subordinate clauses.

In the main-main condition, the Anp instances have interesting properties in com-mon. Four out of the seven Anp cases involved complex NPs in which both competingantecedents belong to the complex NP construction. It turned out that the rankingwe assumed for complex NPs did not always predict the intended referent correctly.For example, in (63), the ranking of the complex NP i simetohi tu k. Avramopulu (Mr.Avramopoulos’s participation) is Avramopulos > Simetohi because Avramopulos (thecurrent mayor of Athens) is animate and ranks higher. The intended referent of the

28 Not surprisingly, chi-square gives a highly significant p < 0.0005.

351


Table 2Reference in main and subordinate clauses.

Ap Anp TotalMain-main 81 (92%) 7 (7%) 88Main-subordinate 55 (51%) 53 (49%) 108

dropped subject in the co-ordinated clause, however, is simetohi. A possible explana-tion for this pattern is to analyze (63) as verb phrase (VP) coordination, in which casethe two VPs share the same subject.

(63) a. Apo afto prokiptei oti [i simetohi]i [tu k. Avramopulu]j

From this concludes that the participation of-the Mr. Avramopulosstis prosehis ekloges epireazi apofasistika tin tihi tis NDat-the next elections affects decisively the fate of-the ND-(name ofpolitical party)From this it is concluded that Mr. Avramopoulos’s participation at thenext elections decisively affects the fate of ND

b. ke [0]i evnoi antistihos to PaSoK.and 0 favors correspondingly the PaSoK-(name of political party)and [it]i favors PaSoK at the same time.

The same phenomenon was observed, however, in cases with no VP coordination,as shown in (64). Again, in this case the anaphoric resolves to koma (‘political party’)and not to Avramopulos, as would be expected. A possible explanation here is that theconcept political party is not “inanimate” in the sense that it denotes a particular groupof people. In this case, it would be an animate possessor, and the ranking would workas expected. In conclusion, although it seems possible that the ranking of complexNPs can be fixed reliably, taking into account special cases of VP coordination andanimacy of collective nouns, the number of instances of such special cases is too smallto draw any definitive conclusions.

(64) a. [To koma]i [tu Avramopulu]i emfanizete se thesi na anadihthithe party of-the Avramopulos appears in position to be-promotedse paragonta pu tha tropopiisi tus orus tu politiku pehnidiu.to factor which will change the terms of-the political game.Avramopoulos’s political party appears to be in a position to getpromoted to a factor that will change the terms of the political game.

b. Me to 14,7% pu pistonete os ‘prothesi psifu’ [0]i katagrafiwith the 14.7% which gets-credited as ‘intention of-vote’ 0 recordsaxiologi apihisi protu kan anadihthun ta politika haraktiristikasignificant appeal before even get-revealed the political characteristicstu.its.With the 14.7% which gets recorded as ‘vote intention’, [it]i records asignificant appeal even before its political characteristics are shown.

The remaining cases of reference to Anp involved complex discourses in whicheither inferrable information was needed or the referent was placed in an adverbial

352


located in the same clause as the anaphoric itself. The following example containsinstances of both cases:

(65) a. Legete oti [o ‘Mihanismos’]i metaferotan stin Arheait-is-said that the ‘Mechanism’ was-being-transfered to-the AncientRomi gia na epidihthi ston KikeronaRome in-order that be-shown to-the KikeronaIt is said that the ‘Mechanism’ was being transfered to Ancient Romein order to be shown to Cicero

b. ala to plio vithistike exo apo ta Kithira.but the boat sank outside of the Kithira.but the boat sank off (the coast of) Kithira.

c. To navagio entopistike stis arhes tu eonathe shipwreck was-located at-the beginning of the centuryThe shipwreck was found at the beginning of the century

d. ke meta tin anelkisi mathimatiki ke arheologi [ton]i

and after the hoisting mathematicians and archaeologists itanasinthesan.they-reconstructed.and after the hoisting, mathematicians and archaeologistsreconstructed iti.

The pronoun in (65d) resolves to Mihanismos in (65a). The entity Mihanismos is evokedmuch more recently as an inferrable entity in (65d), however—the hoisting of theMechanism—and it appears in the same clause as the anaphoric itself. Such complexcases are extremely rare and generally very hard to resolve with a structure-basedalgorithm.

To complete the analysis of the data, we further broke down the distribution ofreference to Ap and Anp for each subordinate clause. The results are shown in Table 3.Chi-square shows no significant differences among the three types of subordinateclauses (p < 0.182).

These results indicate that the focusing preferences of the connectives do not bythemselves predict the interpretation of the anaphoric expressions. They are, however,consistent with Stevenson et al.’s (2000) conclusions that the effect of the connectiveon the interpretation of pronominals depends on the event structure of the precedingclause, either reinforcing or reducing the effect of the verb focusing projections. Lackof correlations between subordinate type and anaphora resolution is not surprising,since the data included various types of verbs.

Table 3Distribution of Ap/Anp.

Ap Anp TotalMain-when(otan) 23 (48%) 25 (52%) 48Main-because(yati) 6 (35%) 11 (65%) 17Main-so that(oste) 26 (60%) 17 (40%) 43

353


5. General Discussion

The interpretation of anaphoric expressions in natural language processing is not atrivial problem. Extensive research in past years has made significant contributionsto our understanding of the phenomenon, and a considerable number of theoreti-cally motivated and/or corpus-based anaphora resolution algorithms have been builtwith more or less success. The task remains a challenge, however, and the slowrate of improvement in the performance of anaphora resolution systems is somewhatalarming.

The detailed review of the literature provided in Section 2 revealed that many ofthe complications and inconsistencies in anaphora resolution start when algorithmsare faced with anaphoric elements in complex sentences. In particular, we saw thatthe interpretation of anaphoric expressions in certain types of clauses would defy anyalgorithm based on registers of NPs and a uniform lookup mechanism.

The main contribution of this article is precisely the distinction and specificationof two systems that determine preferences for anaphoric interpretation. Contra earlierviews on the status of subordinate clauses, we argued that subordinate clauses donot constitute independent processing units. In fact, subordinate clauses can be seenas filling up extended argument positions required by the predicate of the matrixclause, and, in this respect, intrasentential relationships that hold between predicatesand participating entities should be expected to be closely determined on semanticgrounds. We identified the boundaries of the basic discourse units with the boundariesof the unit containing a matrix clause and all its dependent clauses and suggested thatanaphoric interpretation within this unit is determined semantically by the focusingproperties of the verbs and connectives.

On the other hand, topic continuity, as evaluated in the centering model, requiresrather arbitrary specifications of salience to facilitate discourse processing and effi-cient integration of meaning to the previous discourse. Discourses grow enormousvery quickly. Unrestricted semantic representations and the resulting inferencing loadimposed by exploding semantic computations would considerably slow down dis-course processing (Kohlhase and Koller 2000). The notion of salience, in the senseof centering (Joshi and Kuhn 1979), is arguably crucial for efficient processing notonly for natural language processing systems, but also for humans. Topic continuitytherefore is evaluated using a salience mechanism operating across processing units,and we showed that this mechanism is structural and best defined in centering terms.We then argued that anaphoric reference that spans across units is also determinedstructurally.

Regarding centering-based anaphora resolution algorithms, which seem the bestcandidates for resolving anaphora across units, a few technical problems were dis-cussed in Section 2. We suggested, however, that these problems can be fixed easilyand proposed that the algorithm should select as the preferred antecedent the highest-ranked entity in the previous unit. This modification is, in fact, consistent with cen-tering’s pronoun rule and at the same time does not rely on the assumption that textis maximally coherent.

The corpus-based study reported in Section 5.2 was designed to test the hypothesisthat two mechanisms are indeed at work and also to evaluate the strengths of themodified centering-based algorithm for resolving anaphoric reference across units.The results were robust despite the moderate sample size, prescribing a route for anumber of future projects in this direction, the most challenging of which will probablybe understanding the structural and semantic properties of subordination and its rolein the organization, representation, and structure of discourse.

354


AcknowledgmentsI am most grateful to Ellen Prince andAravind Joshi as well as to Robin Clark,Barbara Grosz, Michael Kohlhase,Alexander Koller, and Alistair Knott forvery useful and stimulating discussions. Iwould also like to thank Felicia Hurewitz,Jesse Snedeker, and John Trueswell for theirgenerous help in the design and analysis ofthe experiment as well as the othermembers of the Gleitmans’ CHEESEseminar in the Psychology Department,University of Pennsylvania, for theircontinuous support and valuable feedback.Finally, I am grateful to three anonymousreviewers, whose thoughtful andconstructive criticism helped me improvethe quality of this article significantly. Theresearch on which this article is based wassupported by the Institute of Research inCognitive Science, grant no. NSF-SBR8920230, University of Pennsylvania.

ReferencesBrennan, Susan, Marilyn Walker-Friedman,

and Carl Pollard. 1987. “A centeringapproach to pronouns.” In Proceedings ofthe 25th Annual Meeting of the Association forComputational Linguistics, pages 155–162,Stanford, California.

Carden, Guy. 1982. Backwards anaphora indiscourse context. Journal of Linguistics18:361–387.

Danes, Frantisek. 1974. “Functional sentenceperspective and the organization of thetext.” In F. Danes, editor, Papers onFunctional Sentence Perspective. Academia,Prague, pages 106–128.

Di Eugenio, Barbara. 1990. “Centeringtheory and the Italian pronominalsystem.” In Proceedings of the 13thInternational Conference on ComputationalLinguistics (COLING 90), pages 270–275,Helsinki.

Di Eugenio, Barbara. 1998. “Centering inItalian.” In M. Walker, A. Joshi, andE. Prince, editors, Centering Theory inDiscourse. Clarendon Press, Oxford, pages115–137.

Dimitriadis, Alexis. 1996. “When pro-droplanguages don’t: Overt pronominalsubjects and pragmatic inference.” InProceedings of the 32nd Meeting of the ChicagoLinguistics Society, Chicago, pages 33–47.

Grosz, Barbara, Aravind Joshi, and ScottWeinstein. 1983. “Providing a unifiedaccount of definite noun phrases indiscourse.” In Proceedings of the 21stAnnual Meeting of the Association forComputational Linguistics, pages 44–50,

MIT Press, Cambridge, Massachusetts.Grosz, Barbara and Candace Sidner. 1986.

Attentions, intentions and the structure ofdiscourse. Computational Linguistics12(3):175–204.

Hobbs, Jerry. 1978. Resolving pronounreferences. Lingua 44:311–338.

Hudson-D’Zmura, Susan and MichaelTanenhaus. 1998. “Assigning antecedentsto ambiguous pronouns: The role of thecenter of attention as a defaultassignment.” In M. Walker, A. Joshi, andE. Prince, editors, Centering Theory inDiscourse. Clarendon Press, Oxford, pages199–226.

Hurewitz, Felicia. 1998. “A quantitative lookat discourse coherence.” In M. Walker,A. Joshi, and E. Prince, editors, CenteringTheory in Discourse. Clarendon Press,Oxford, pages 273–291.

Joshi, Aravind and Steven Kuhn. 1979.“Centered logic: The role of entitycentered sentence representation innatural language inferencing.” In SixthInternational Joint Conference on ArtificialIntelligence, pages 435–439, Tokyo.

Kameyama, Megumi. 1993. “Intrasententialcentering.” In Proceedings of the Workshopon Centering, University of Pennsylvania,Philadelphia.

Kameyama, Megumi. 1998. “Intrasententialcentering: A case study.” In M. Walker,A. Joshi, and E. Prince, editors, CenteringTheory in Discourse. Clarendon Press,Oxford, pages 89–112.

Kohlhase, Michael and Alexander Koller.2000. “Towards a tableaux machine forlanguage understanding.” In Proceedingsof the Second Workshop on Inference inComputational Semantics (ICOS-2), pages57–88, Dagstuhl, Germany.

Kuno, Susumu. 1972. Functional sentenceperspective: A case study from Japaneseand English. Linguistic Inquiry 3:269–320.

Lappin, Shalom and Herbert Leass. 1994.An algorithm for pronominal anaphoraresolution. Computational Linguistics20(4):535–561.

McCord, Michael. 1990. “Slot grammar: Asystem for simpler construction ofpractical natural language grammars.” InR. Studer, editor, Natural Language andLogic: International Scientific Symposium.Lecture Notes in Computer Science.Berlin, Springer Verlag, pages 118–145.

Miltsakaki, Eleni. 1999. “Locating topics intext processing.” In ComputationalLinguistics in the Netherlands: Selected Papersfrom the Tenth CLIN Meeting (CLIN ’99),pages 127–138, Utrecht, Netherlands.

Miltsakaki, Eleni. 2000. Attention Structure

355


in Discourse: A Cross-LinguisticInvestigation. Dissertation topic proposal,University of Pennsylvania, Philadelphia.

Miltsakaki, Eleni. 2001. “Centering inGreek.” In Proceedings of the 15thInternational Symposium on Theoretical andApplied Linguistics, Thessaloniki, Greece.

Miltsakaki, Eleni and Karen Kukich. 2000a.“Automated evaluation of coherence instudent essays.” In Proceedings of theWorkshop on Language Rescources and Toolsin Educational Applications, LREC 2000,pages 7–14, Athens.

Miltsakaki, Eleni and Karen Kukich. 2000b.“The role of centering theory’s rough shiftin the teaching and evaluation of writingskills.” In Proceedings of the 38th AnnualMeeting of the Association for ComputationalLinguistics (ACL 2000), pages 408–415,Hong Kong.

Prasad, Rashmi and Michael Strube. 2000.“Pronoun resolution in Hindi.” InA. Williams and E. Kaiser, editors,Working Papers in Linguistics, volume 6.University of Pennsylvania, pages189–208.

Quirk, Randolph, Sidney Greenbaum,Geoffrey Leech, and Jan Svartvik. 1972. AGrammar of Contemporary English.Longman, London.

Rambow, Owen. 1993. Pragmatic aspects ofscrambling and topicalization in German.In Workshop on Centering Theory inNaturally Occurring Discourse, Institute ofResearch in Cognitive Science, Universityof Pennslylvania.

Stevenson, Rosemary, Rosalind Crawley,and David Kleinman. 1994. Thematicroles, focusing and the representation ofevents. Language and Cognitive Processes9:519–548.

Stevenson, Rosemary, Alistair Knott, JonOberlander, and Sharon McDonald. 2000.Interpreting pronouns and connectives:Interactions among focusing, thematicroles and coherence relations. Languageand Cognitive Processes 15(3):225–262.

Strube, Michael. 1998. “Never look back: Analternative to centering.” In Proceedings ofthe 17th International Conference onComputational Linguistics and the 36thAnnual Meeting of the Association forComputational Linguistics (COLING-ACL’98), pages 1251–1257, Montreal.

Strube, Michael and Udo Hahn. 1996.“Functional centering.” In Proceedings ofthe 34th Annual Conference of the Associationfor Computational Linguistics (ACL ’96),pages 270–277, Santa Cruz.

Strube, Michael and Udo Hahn. 1999.Functional centering: Grounding

referential coherence in informationstructure. Computational Linguistics25(3):309–344.

Suri, Linda and Kathleen McCoy. 1994.RAFT/RAPR and centering: Acomparison and discussion of problemsrelated to processing complex sentences.Computational Linguistics 20(2):301–317.

Suri, Linda, Kathleen McCoy, and JonDeCristofaro. 1999. A methodology forextending focusing frameworks.Computational Linguistics 25(2):173–194.

Tanaka, Izumi. 2000. “Cataphoric personalpronouns in English news reportage.” InProceedings of the Third InternationalDiscourse Anaphora and Anaphor ResolutionConference, DAARC 2000, volume 12,pages 108–117, Lancaster, England.

Turan, Umit. 1998. “Rankingforward-looking centers in Turkish:Universal and language specificproperties.” In M. Walker, A. Joshi, andE. Prince, editors, Centering Theory inDiscourse. Clarendon Press, Oxford, pages139–160.

van Hoek, Karen. 1997. Anaphora andConceptual Structure. University ofChicago Press, Chicago.

Walker, Marilyn, Masayo Iida, and SharonCote. 1994. Japanese discourse and theprocess of centering. ComputationalLinguistics 20(2):193–233.

Walker, Marilyn and Ellen Prince. 1996. “Abilateral approach to givenness: Ahearer-status algorithm and a centeringalgorithm.” In T. Fretheim and J. Gundel,editors, Reference and Referent Accessibility.John Benjamins, Amsterdam, pages291–306.

Webber, Bonnie and Aravind Joshi. 1998.“Anchoring a lexicalized tree adjoininggrammar for discourse.” In ACL/COLINGWorkshop on Discourse Relations andDiscourse Markers, pages 8–92, Montreal,Quebec, Canada.

Webber, Bonnie, Alistair Knott, MatthewStone, and Aravind Joshi. 1999a.“Discourse relations: A structural andpresuppositional account usinglexicalized TAG.” In Proceedings of the 37thAnnual Meeting of the Association forComputational Linguistics, pages 41–48,College Park, Maryland.

Webber, Bonnie, Alistair Knott, MatthewStone, and Aravind Joshi. 1999b. “Whatare little texts made of? A structural andpresuppositional account usinglexicalized TAG.” In Proceedings of theInternational Workshop on Levels ofRepresentation in Discourse (LORID ’99),pages 145–149, Edinburgh.

toward an aposynthesis of topic continuity and ... · toward an aposynthesis of topic continuity...

Documents